CVE-2026-108957 Details

A vulnerability allowing arbitrary code execution has been identified in TinyLLaVA TinyLLaVA_Factory versions 1.0.0 prior to 1abc6ac5e76a68d0fd087a071b56da968e69ca9b. The issue arises in the TinyLlavaConfig component, specifically within the _load_text_config function of the file tinyllava/model/configuration_tinyllava.py. The vulnerability is triggered by manipulating the llm_model_name_or_path argument, which leads to code injection. This flaw can be exploited remotely, and the released exploit takes advantage of the hardcoded trust_remote_code=True, allowing execution of attacker-controlled Python code during the configuration loading process.

Users are advised to remove the hardcoded trust_remote_code=True in the TinyLlavaConfig._load_text_config method. Instead, the trust_remote_code flag should be explicitly passed from the caller, defaulting to False. Additionally, consider rejecting local relative paths in llm_model_name_or_path unless trust_remote_code has been explicitly enabled by the user.